fix(codegen): gate closure-local-call implicit-this reset on callee reading dynamic this

[proggeramlug]

Summary

Fixes #5030 — unbreaks the required compiler-output-regression check, red on main since #5004 (verified by adjacent-commit runs: gate green at 50c1a64 / #5003, red at f991e2e / #5004 with the exact failure signature from CI).

Root cause

#5004 added an OrdinaryCallBindThis reset (js_implicit_this_set(undefined) + restore) around receiverless calls. The FuncRef path correctly gates it on funcs_reading_dynamic_this ("ordinary helper calls pay nothing"), but try_lower_closure_typed_local_call (early_branches.rs) emits the pair unconditionally. In the h1_buffer_alias_negative fixture, closureCapture()'s hot loop calls a this-free arrow (read(i)) — the per-iteration js_implicit_this_set pair trips hot_loops_no_runtime_calls (it is the only call in the failing report that is not on the workload's allowlist).

Fix

Mirror the FuncRef gating in the closure-typed-local path:

• Direct/guarded path, statically-known callee: emit the reset only when the mapped closure's body reads dynamic this (funcs_reading_dynamic_this, which conservatively includes arrows capturing lexical this).
• Typed-feedback guard fallback: the runtime receiver is some other closure codegen never saw — when the static gate skipped the outer reset, the fallback block now does its own reset/restore around js_closure_callN. (Fallback blocks are not hot-loop blocks, so the gate stays green while runtime/object-model: implement strict and sloppy function this binding #3576 semantics are preserved even for reassigned closure locals.)
• Generic js_closure_callN tail (unknown func id / rest params / arity mismatch): unconditional reset, as before.

Validation

• compiler_output_regression.py suite --suite native-region-proof → all workloads pass (failed_workloads: []), previously h1_buffer_alias_negative failed.
• capture --gate for vectorized_buffer_transform and hir_fact_rewrite → pass.
• python3 -m unittest tests.test_compiler_output_regression → 40/40.
• test-files/test_object_model_semantics_3986.ts (the fix(runtime,hir): object-model semantics — Object boxing, this-binding, var-hoisting (#3986/#3577/#3576/#3575) #5004 semantics test) still prints object-model-semantics: ok under Perry.
• New targeted behavioral check (compiled + ran): (1) closure local statically mapped to a this-free arrow but reassigned at runtime to a this-reading function — guard fallback binds this to undefined/globalThis, no receiver leak; (2) this-reading closure local on the fast path — reset still applied; (3) the CI: compiler-output-regression red on main (h1_buffer_alias_negative native-region-proof workload) #5030 hot-loop helper shape computes correctly.
• cargo test --release -p perry-codegen → green except manifest_param_counts_match_dispatch_table (zlib::deflateRawSync arity drift), which is pre-existing macOS-local and unrelated (my diff touches neither manifest nor dispatch; the test passes in Linux CI on current main).

Related: #5031 fixes the cargo-test red (stream unshift); #5029 (gc_write_barrier_stress) is being bisected separately.

No version bump / changelog per external-PR convention.

[proggeramlug]

Superseded by #5033 (same commit, stacked with the stream fix and the #5029 test gating so all required checks can pass together).